Computing Sensitivities in Reaction Networks using Finite Difference Methods

TL;DR

This paper evaluates numerical methods for computing sensitivities in reaction networks, comparing accuracy and performance, and recommends the two-point and five-point methods for different scenarios.

Contribution

It systematically compares several finite difference methods for sensitivities in reaction networks, providing practical recommendations based on accuracy and speed.

Findings

Richardson's extrapolation is most accurate but slowest.

Two-point method offers a good balance of speed and accuracy.

Five-point method is suitable for rapidly changing derivatives.

Abstract

In this article, we investigate various numerical methods for computing scaled or logarithmic sensitivities of the form $\partial \ln y/\partial \ln x$. The methods tested include One Point, Two Point, Five Point, and the Richardson Extrapolation. The different methods were applied to a variety of mathematical functions as well as a reaction network model. The algorithms were validated by comparing results with known analytical solutions for functions and using the Reder method for computing the sensitivities in reaction networks via the Tellurium package. For evaluation, two aspects were looked at, accuracy and time taken to compute the sensitivities. Of the four methods, Richardson's extrapolation was by far the most accurate but also the slowest in terms of performance. For fast, reasonably accurate estimates, we recommend the two-point method. For most other cases where the derivatives are changing rapidly, the five-point method is a good choice, although it is three times slower than the two-point method. For ultimate accuracy which would apply particularly to very fast changing derivatives the Richardson method is without doubt the best, but it is seven-times slower than the two point method. We do not recommend the one-point method in any circumstance. The Python software that was used in the study with documentation is available at: \url{https://github.com/evanyfyip/SensitivityAnalysis}.